Evolving Benchmarks

Quite often algorithms are compared by running them on a standard optimisation problem. We have shown this can easily be turned on its head and new benchmarks can be automatically designed which prefer one optimiser over another and vice-versa. By studying how both algorithms behave on the new benchmark we can learn about the optimizers strengths and weaknesses. Paper at CEC 2005 (slides).

Evolved benchmark where simple PSO beats differential evolution.

PSO (no constriction) Particle 0 Generation 0-44 0.0003(0.11+x+0.7y)

Trace of one particle in the PSO swarm on landscape evolved to suit Differential Evolution

The above animation shows a single PSO particle. (The other 29 members of the swarm behave similarly.) Note that it oscillates about its own and the global best (coloured dots).
   Adding constriction to the PSO changes the behaviour of the swarm so that it can now readily solve the problem and find the global optimum (animation).

Links

XPS team
My Papers

W.B.Langdon 20 Feb 2006